Olli Karjalainen Nordia
Geographical Publications
Volume 49:1
Statistical modelling of circumpolar permafrost: thermal and geomorphic
sensitivities to climate change and societal implications
to be presented with the permission of the Doctoral Training Committee for Technology and Natural Sciences of the University of Oulu Graduate School
(UniOGS), for public discussion in the lecture hall IT116, on the 13th of March, 2020, at 12 noon.
ACADEMIC DISSERTATION
Nordia
Geographical Publications
Volume 49:1
Statistical modelling of circumpolar permafrost: thermal and geomorphic
sensitivities to climate change and societal implications
Olli Karjalainen
Nordia Geographical Publications Publications of
The Geographical Society of Northern Finland and
Address: Geography Research Unit P.O. Box 3000
FIN-90014 University of Oulu FINLAND
olli.karjalainen@oulu.fi
Editor: Teijo Klemettilä
Nordia Geographical Publications ISBN 978-952-62-2541-8
ISSN 1238-2086
Punamusta Oy Tampere 2020
Geography Research Unit, University of Oulu
Statistical modelling of circumpolar permafrost:
thermal and geomorphic sensitivities to climate change
and societal implications
v
Contents
Abstract vii
Original publications x
Acknowledgements xii
Glossary xiii
1 Introduction 1
2 Study background 5
2.1 Preconditions for terrestrial permafrost 5
2.1.1 Permafrost-climate relationship 6
2.1.2 Climatic sensitivity of permafrost 7
2.1.3 Terrain properties 8
2.2 Permafrost landforms 10
2.3 Human activity and permafrost 12
3 Study area 13
4 Materials and methods 15
4.1 Response data 15
4.1.1 Observations of mean annual ground temperature and
active-layer thickness 15
4.1.2 Permafrost landform observations 16
4.2 Geospatial data 17
4.2.1 Current and future climates 20
4.2.2 Terrain properties 20
4.2.3 Infrastructure data 21
4.3 Statistical modelling 22
4.3.1 Geohazard assessments 24
4.3.2 Model evaluation 24
vi
5 Results and discussion 27
5.1 Circumpolar controls of permafrost (RQ 1) 29 5.2 Permafrost in changing climates (RQ 2) 32 5.2.1 Permafrost extent and active-layer thickness dynamics 32 5.2.2 Potential environmental spaces for permafrost
landform occurrence 34
5.3 Geohazards related to near-surface permafrost
degradation (RQ 3) 38
5.4 Infrastructure at risk (RQ 4) 40
5.5 Methodological considerations (RQ 5) 42
5.5.1 Spatial modelling in time 43
6 Future study needs 45
7 Conclusions 47
References 49
Original publications
vii
Abstract
Statistical modelling of circumpolar permafrost: thermal and geomorphic sen- sitivities to climate change and societal implications
Karjalainen, Olli, Geography Research Unit, University of Oulu, 2020
Keywords: permafrost, ground temperature, active layer, geospatial data, statistical modelling, GIS, circumpolar, Arctic, pingo, ice-wedge polygon, rock glacier, geomorphology, geohazard, permafrost thaw, infrastructure risk, climate change.
One-fourth of the land area in the Northern Hemisphere is affected by perennially frozen ground, known as permafrost. The thermal conditions of permafrost govern complex geo- and ecosystems and provide support for Arctic cities and transportation infrastructure.
Permafrost, however, is not permanent. Rather it is sensitive to the warming climate and human-induced disturbances. Recently, rapid degradation of permafrost landscapes has been observed across the Arctic. In addition to the local implications for the hydro- ecology, geo- and biodiversity and ground stability, permafrost degradation can affect the global climate through biogeochemical feedbacks. Ongoing changes to Arctic permafrost systems may have environmental and socio-economic repercussions on national and international scales.
The main aims of this thesis were to first examine how environmental conditions control the thermal and geomorphic permafrost characteristics on a circumpolar scale. Next, the sensitivity of permafrost to 21st century climate change was assessed. Lastly, high-resolution geohazard maps were produced and used to quantify the amount of infrastructure potentially at risk from thawing near-surface permafrost across the Northern Hemisphere. The thesis utilized statistical ensemble modelling techniques and geospatial datasets combined with comprehensive circumpolar observational datasets.
Based on the results, the studied permafrost characteristics were strongly controlled by and sensitive to current and future climatic conditions. The air temperature and rainfall had the most prominent contributions, while the effects of local terrain properties on a circumpolar scale were often found to be small. By the mid-century, the extent of near- surface permafrost may decrease by 34–47% depending on human-induced greenhouse gas emissions. Suitable areas for permafrost landform occurrence will similarly shrink, including regions of cold continuous permafrost. Quantifications of the infrastructure at risk indicated that around 70% of the studied engineering elements and four million people were located in areas of projected near-surface permafrost thaw. Moreover, one-third of all the infrastructure elements and nearly a million people situated in regions with high potential for permafrost degradation-related damage to the built environment.
viii
In conclusion, circumpolar permafrost was projected to show extensive regionally distinct sensitivities to the ongoing climate change. Although most of the thermal and geomorphic impacts were projected to occur by the mid-century regardless of the climate- change scenario, it is argued that mitigating climate change could reduce the potential consequences for natural and human systems. In order to achieve a higher applicability of circumpolar-scale analyses on local scales, further developments in the availability and quality of global observational and geospatial data are required. Moreover, it is proposed that the pronounced nonlinearity found between climatic conditions and the studied permafrost characteristics should be carefully considered in future assessments.
ix
Supervisors
Professor Jan Hjort Geography Research Unit University of Oulu, Finland Professor Miska Luoto
Department of Geosciences and Geography University of Helsinki, Finland
Pre-examiners
Professor Stephan Gruber
Department of Geography and Environmental Studies Carleton University, Canada
Professor Nikolay I. Shiklomanov Department of Geography
The George Washington University, U.S.A.
Official opponent
Professor Martin Hoelzle Department of Geosciences University of Fribourg, Switzerland
x
Original publications
I Karjalainen, O., Luoto, M., Aalto, J. & J. Hjort (2019). New insights into the environmental factors controlling the ground thermal regime across the Northern Hemisphere: a comparison between permafrost and non-permafrost areas. The Cryosphere, 13, 693–707.
II Karjalainen, O., Luoto, M., Aalto, J., Etzelmüller, B., Grosse, G., Jones, B.M., Lilleøren, K.S. & J. Hjort. High potential for loss of iconic permafrost landforms in a changing climate (Submitted manuscript).
III Karjalainen, O., Aalto, J., Luoto, M., Westermann, S., Romanovsky, V.E., Nelson, F.E., Etzelmüller, B. & J. Hjort (2019). Circumpolar permafrost maps and geohazard indices for near-future infrastructure risk assessments. Scientific Data, 6, 190037.
IV Hjort, J., Karjalainen, O., Aalto, J., Westermann, S., Romanovsky, V.E., Nelson, F.E., Etzelmüller, B. & M. Luoto (2018). Degrading permafrost puts Arctic infrastructure at risk by mid-century. Nature Communications, 9, 5147.
Published articles I, III and IV are licensed under the Creative Commons Attribution 4.0 International License.
xi Original idea Manuscript
preparation Data compilation and mining
Statistical
analysis GIS &
geospatial analysis Paper I OK, ML, JH OK, JA, JH,
ML OK, JA OK, ML,
JA, JH OK, JA Paper II OK, ML, JH OK, JH,
ML, JA, BE, GG, BMJ, KSL
OK, JH, BMJ, GG, BE, KSL
OK, JH,
ML, JA OK, JH, JA
Paper III JH, OK, ML, JA, VER, FEN, SW, BE
OK, JA, JH, ML, SW, FEN, VER, BE
OK, JH,
ML, JA JA, OK,
JH, ML OK, JA
Paper IV ML, JH JH, OK, JA, ML, SW, FEN, VER, BE
OK JH, ML,
JA JA, JH,
ML, OK OK, JA
Author contributions
Aalto, Juha = JA
Etzelmüller, Bernd = BE Grosse, Guido = GG Hjort, Jan = JH
Jones, Benjamin M. = BMJ Karjalainen, Olli = OK Lilleøren, Karianne S. = KSL Luoto, Miska = ML
Nelson, Frederick E. = FEN Romanovsky, Vladimir E. = VER Westermann, Sebastian = SW
xii
Acknowledgements
Doctoral studies have been most exciting times. Before taking this path, I could not have fathomed the huge challenges and triumphant highs that it would lay before me. I am deeply grateful to my principal supervisor Jan Hjort for being an excellent scientific mentor and for making it possible for me to take such a fascinating start on an academic career. Next, I wish to thank my supervisor Miska Luoto (University of Helsinki) for his uncompromising and supportive feedback. You helped me to seek the best I had to give.
The same goes for the fourth member of our ‘Team Finland’ from the past years, Juha Aalto, for painstakingly helping me to improve my work.
The research presented in the thesis involved many key persons from around the Northern Hemisphere permafrost and non-permafrost regions. Special thanks go to Bernd Etzelmüller, for facilitating my research visit at the University of Oslo and for the very insightful discussions on rock glaciers and more. Furthermore, I am in debt for the knowledge and efforts that all the co-authors put in; thank you, Guido Grosse, Benjamin M. Jones, Karianne S. Lilleøren, Frederick E. Nelson, Vladimir E. Romanovsky and Sebastian Westermann. I am happy to have done research with you and hope to see it continue. I am very grateful to Martin Hoelzle for acting as my opponent in the defence.
I also wish to thank the pre-examiners, Stephan Gruber and Nikolay I. Shiklomanov, for their constructive criticism on the summary section.
My dear colleagues at the whole Geography Research Unit, be acknowledged that you have been a joyful bunch to work with. With you the future of geography in Oulu as a discipline and an academic identity is in good hands. I am especially happy for the diverse activities we had with the Physical Geography Research Group and the Geographical Society of Northern Finland. My follow-up group members, Janne Alahuhta and Jarmo Rusanen, thank you for your time and support. Additionally, I wish to thank the head of the unit Jarkko Saarinen and acknowledge funding from two Academy of Finland projects (285040 and 315519) for facilitating me with a top-class environment for conducting scientific work. Further thanks go to the people of the University of Oulu Graduate School for supporting my educational development and for granting travel money.
Above all, I would like to express my gratitude to my loved Tiina-Mari and Iisa-Maija.
You were always there for me making my life complete. Lastly, deep appreciation to my parents and sisters for all their support and keen interest in my studies over the years.
Oulu, January 2020 Olli Karjalainen
xiii
Glossary
Active layer = Seasonally thawed layer on top of permafrost.
Depth of zero annual amplitude (DZAA) = The depth at which the annual variation of ground temperature is < 0.1 °C.
Digital Elevation Model (DEM) = DEMs are digital representations of elevation on the Earth’s surface. They can be used to compute topography parameters.
Ensemble modelling = The notion of an ensemble implies that combining multiple predictions with different statistical assumptions can yield an improved prediction and reduce uncertainty.
Equilibrium assumption = Here, equilibrium modelling assumes that a permafrost property is in thermal/geomorphic balance with surrounding environmental conditions (e.g. climate).
Generalized Additive Model (GAM) = A semiparametric extension of GLM, wherein model terms can be fitted nonparametrically using a smoothing function.
Generalized Boosting Method (GBM) = An ensemble learning method, in which multiple decision trees are fitted using boosting, i.e. the fits of previous fitted trees are considered sequentially to improve the accuracy of the final model.
Generalized Linear Model (GLM) = An extension of an ordinary linear regression, in which a combination of explanatory variables is related to a response variable by a link function.
Geohazard index = The geohazard index is a statistically formulated spatial representation of a geohazard, i.e. a potentially harmful geological or environmental condition.
Geospatial data = Data on a feature or phenomena assigned with location properties.
Ground ice = Frozen water content in the ground that can occur in various forms (e.g.
segregated ice, massive ice or wedge ice).
Infrastructure = Engineering structures with permanent foundations.
Latent-heat exchange = The delay in phase change caused by the demand of extra energy needed to melt ice or freeze water.
Permafrost = Ground material, in which temperature stays at or below 0 °C for at least two consecutive years.
xiv
Permafrost degradation = Considered to occur when the permafrost temperature or the thickness of the active layer increases, or if permafrost thaws.
Permafrost landform = Permafrost landforms are geomorphic manifestations of ground ice processes.
Random Forest (RF) = An ensemble learning method that combines outcomes from multiple decision trees for an improved prediction. In contrast to GBM, each tree is fully grown, i.e. using all instances in randomly chosen subsets of predictors.
Representative Concentration Pathway (RCP) = RCPs depict atmospheric greenhouse gas concentration trajectories.
Statistical distribution modelling = Predictive modelling of the spatial distribution of a physical property based on a statistical model. In the modelling, observations of the property are statistically related to environmental conditions.
Thermokarst = Thermokarst is used to represent the process where the ground surface collapses (forming e.g. depressions, thaw slumps or thermokarst lakes) as the ground ice melts.
1
Cold climates host a unique set of Earth surface processes behind the landscape evolution. Permafrost, or perennially frozen ground, is a thermally defined phenomenon that affects almost one-quarter of the Northern Hemisphere landmasses (Zhang et al.
1999). Soil or bedrock has permafrost if its temperature remains at or below 0 °C for at least two consecutive years (Permafrost Subcommittee 1988). Permafrost occurs in a delicate balance with climatic conditions, and thus reflects spatial patterns and dynamics in climatic conditions on seasonal to millennial time scales (Washburn 1980; Hinzman et al. 2005; Throop et al. 2012; Chadburn et al. 2017). The dynamic nature of the ground thermal regime suggests that permafrost is not permanent, which underlines the urgency of studying the state of permafrost in the rapidly changing Arctic climate.
A general increase in permafrost temperatures has been observed across the Northern Hemisphere during the last few decades (e.g. AMAP 2017; Romanovsky et al. 2018;
Biskaborn et al. 2019; Meredith et al. 2019). The primary reason behind the observed trends is the increasing air temperature, but in some regions the snow thickness also exerts a definitive control on permafrost (Biskaborn et al. 2019). Consequently, near- surface permafrost in many regions has thawed (e.g. Smith et al. 2010; Nicolsky et al.
2017; Romanovsky et al. 2010, 2017). This development is likely to continue as the Arctic is warming more rapidly than the Earth on average in the twenty-first century (Hoegh- Guldberg et al. 2018; Schuur & Mack 2018). Alongside warming mean annual ground temperatures (MAGT), the depth of the seasonally thawed layer on top of permafrost, known as the active-layer thickness (ALT), has increased in many permafrost regions (Park et al. 2013; Luo et al. 2016; AMAP 2017; Romanovsky et al. 2018). In this thesis, MAGT and ALT are used to characterize the thermal state of permafrost, implying that increasing MAGT or ALT are indicative of permafrost degradation.
The climate-induced thermal dynamics of permafrost have not been uniform. In general, cold permafrost has been warming more rapidly than warmer permafrost (Smith et al. 2010; Romanovsky et al. 2010, 2017; Biskaborn et al. 2019). In contrast, the most momentous ALT increase has occurred in warm permafrost (Luo et al. 2016; AMAP 2017). The degradation of near-surface permafrost has the most dramatic consequences in areas with high ground ice content (Haeberli 1992; Rowland et al. 2010; Streletskiy &
Shiklomanov 2016; Farquharson et al. 2019). When ice-rich permafrost thaws, thermokarst occurs and the ground may subside or become vulnerable to mass movements (Jorgenson et al. 2006; Kokelj & Jorgenson 2013; Schuur et al. 2015). Extensive and accelerating permafrost degradation has been observed recently across the Arctic (Kokelj et al. 2015;
Liljedahl et al. 2016; Jorgenson & Grosse 2016; Farquharson et al. 2019; Lewkowicz & Way 2019) including the Tibetan Plateau (Yang et al. 2010; Ran et al. 2018). This development affects local geomorphological, hydrological and ecological conditions (Liljedahl et al.
2016; Brighenti et al. 2018; Schuur & Mack 2018), but also poses a threat to human
1 Introduction
2 3 constructions on ice-rich permafrost (Vincent et al. 2017; O’Neill et al. 2019; Turetsky
et al. 2019) in lowland (Mackay 1972) and in mountain environments (Haeberli 1992;
Beniston et al. 2018).
The impacts of permafrost degradation are by no means limited to the immediate high-latitude and altitude regions but affect the entire Earth through greenhouse gas and surface albedo feedback, for example (Schuur & Mack 2018; Moon et al. 2019; Turetsky et al. 2019). Bartsch et al. (2016) stress that the links between rapidly changing Arctic and the global climate system necessitate addressing the Arctic in its entirety rather than just regionally. Moreover, research on broad scales is vital because the circumpolar ground thermal regime may have different environmental controls than those described on site, local or regional scales (Riseborough et al. 2008; Grosse et al. 2016). Knight and Harrison (2013) argue that the responses of Earth surface systems to climate change remain poorly understood, notwithstanding that their functioning encompasses critical water and soil resources, ecosystem services, and biogeochemical climate feedback mechanisms.
Moreover, despite abundant knowledge on the current observed permafrost dynamics, the effects and associated feedbacks on the geomorphology, ecosystems and infrastructure on local to global scales are still unclear (Grosse et al. 2011, 2016; Nicolsky et al. 2017;
Oliva et al. 2018) and are limited by often coarse spatio-temporal resolutions of broad- scale analyses and climate-change projections (Riseborough et al. 2008; Etzelmüller 2013;
Romanovsky et al. 2017). The statistical modelling approach applied in this thesis allows for utilizing high resolution geospatial data on environmental conditions and infrastructure elements, and thereby more explicitly than before, predicting local variations in permafrost characteristics and geohazards on a circumpolar scale.
Improved knowledge on the effects of climate change on permafrost regions is essential to facilitate adaptation measures for the likely significant impacts on natural and human systems in the Arctic (Hinzman et al. 2005; Oliva & Fritz 2018; Czekirda et al. 2019). Recent years have yielded a growing number of high-resolution circumpolar- to global-scale (hereafter broad-scale) assessments of the permafrost extent or thermal state (e.g. Gruber 2012; Park et al. 2016; Chadburn et al. 2017; Kroisleitner et al. 2018;
Tao et al. 2018), but analyses using circumpolar-wise comprehensive observational datasets have remained relatively scarce until recently (e.g. Aalto et al. 2018a; Peng et al.
2018; Biskaborn et al. 2019; Obu et al. 2019). Geomorphological landforms which are unique to permafrost regions remain particularly understudied on broad scales. Permafrost landforms are physical manifestations of near-surface ground ice dynamics and they are thus sensitive to permafrost warming and thickening of the active layer (Michel 2011;
Jorgenson et al. 2015).
This thesis aims to produce improved knowledge on the current state of the Northern Hemisphere permafrost and to assess the natural and societal implications of its projected near-future change. More precisely, the following research questions (RQ 1–5) are answered:
3
RQ 1: What are the key circumpolar environmental controls of thermal (MAGT and ALT) and geomorphic (pingos, ice-wedge polygons and rock glaciers) permafrost characteristics (Papers I and II)?
RQ 2: What are the thermal and geomorphic impacts of climate change on permafrost environments (Papers II and III)?
RQ 3: How and where does near-surface permafrost degradation pose hazards to Arctic communities and infrastructure (Papers III and IV)?
RQ 4: What is the magnitude of the potential near-surface permafrost degradation- related damage to infrastructure (Paper IV)?
RQ 5: What are the potential contributions of the used geospatial data-based statistical analyses to permafrost science?
To address these questions, this thesis examines the spatial variation of permafrost characteristics across the Northern Hemisphere by relating observational datasets on MAGT, ALT and permafrost landforms to globally comprehensive geospatial data on environmental conditions at 30 arc-second (< 1 km2) spatial resolution. Climate-change scenarios are then used to assess the future thermal and geomorphic change in permafrost landscapes. The information gained is ultimately applied in assessments of potential societal climate change-induced consequences of degrading near-surface permafrost.
Specifically, spatial datasets are employed to formulate geohazard indices and to quantify which infrastructure elements, natural resource extraction areas, or human settlements and populations are located in areas of permafrost thaw-related geohazards. The analyses are conducted in a statistical spatial modelling framework applying regression and machine learning-based techniques and ensemble modelling.
5
In this section, I present a literature-based conceptualization of the themes covered in the thesis and disclose relevant research gaps. First, the relationship between climate and permafrost characteristics (MAGT and ALT) are discussed. Second, the climatic sensitivity of permafrost is discussed in the context of climate change. Third, a brief review of the terrain properties (topography, soil properties and vegetation) mediating the permafrost-climate relationship on local spatial scales follows. Fourth, I describe the main characteristics of the studied permafrost landforms; pingos, ice-wedge polygons and rock glaciers. Finally, the human aspects of the impacts of degrading permafrost are problematized.
2.1 Preconditions for terrestrial permafrost
Permafrost forms when the annual net radiation balance is negative, i.e. the incoming solar radiation-derived energy affecting the ground is smaller than the outward heat flux from the Earth’s surface (Lachenbruch & Marshall 1969; Péwé 1979; Nicolsky & Romanovsky 2018). Subsurface temperatures are closely connected to the atmospheric conditions affected by diurnal, seasonal, annual, decadal and millennial oscillations (Bodri & Cermak 2007; Harris et al. 2009). As the surface temperature signal propagates downwards, it affects the ground, attenuating progressively with depth (Huang et al. 2000; Bartlett et al.
2004). The amplitude of the annual temperature variation at a given depth depends on the thermal diffusivity of the ground (e.g. Bartlett et al. 2004; French 2007; Smith et al.
2010). The depth at which the annual variation of ground temperature is less than 0.1 °C is considered the depth of zero annual amplitude (DZAA, Fig. 1, Péwé 1979).
2 Study background
6 7 2.1.1 Permafrost-climate relationship
The ground thermal regime is primarily controlled by mean annual air temperature (Smith
& Riseborough 2002; Callaghan et al. 2011; Streletskiy et al. 2015). In general, warmer air temperatures result in a higher MAGT (Throop et al. 2012; Romanovsky et al. 2017) and greater ALT (Hinkel & Nelson 2003; Bonnaventure & Lamoureux 2013), but the local topography, ecosystem and soil conditions notably mediate the effect (Shur & Jorgenson 2007; Etzelmüller 2013; Aalto et al. 2018a, b). In addition, seasonal asymmetries in the air temperature-permafrost linkage are identifiable. Given the equal magnitude of change in winter and summer air temperatures, those during the winter have been suggested to exert a more direct influence on permafrost temperature (Smith & Riseborough 1996;
Etzelmüller et al. 2011; Jones et al. 2016). ALT is essentially dependent on summer
Figure 1. A simplified conceptual presentation of the permafrost thermal regime and the affecting factors.
The black curves depict the hypothetical distribution of winter minimum and summer maximum ground temperatures (T) as a function of depth. The depth at which seasonal variations in ground temperatures fall to below < 0.1 °C is called the depth of zero annual amplitude. The hatched curve represents the mean annual ground temperature (MAGT).
7
temperatures, and a higher annual temperature attributed to a warmer winter typically has a more negligible impact (Oelke et al. 2003; Melnikov et al. 2004; Zhang et al. 2005;
Luo et al. 2016). Warmer winters can however affect the ALT through changing snow conditions and subsequent changes in hydrology and vegetation, for example (Park et al.
2013; Atchley et al. 2016).
The amount and seasonality of precipitation are controlled by the regional climate and further characterized by the local topography (Gruber et al. 2017). Rainfall affects the ground thermal regime by controlling both conductive and convective heat fluxes (Zhang et al. 2001; Westermann et al. 2010, 2011; Marmy et al. 2013; Slater & Lawrence 2013).
For example, Melnikov et al. (2004) argued that heavy summer rains cause warming and a greater ALT, but that in the autumn rain can cool the ground and suppress the thickening of the active layer also in the next thawing season due to the increased ice content at the base of the active layer. Kokelj et al. (2015) stressed that increases in rainfall are anticipated to have a significant impact on permafrost geomorphology.
Snow acts as an insulating agent that increases temperature differences between the air and ground (e.g. Zhang et al. 1996, 2001; Stieglitz et al. 2003; Slater et al. 2017). This is because snow has a low thermal diffusivity, and thus it can mute some of the effects of low air temperatures (Bartlett et al. 2004). The degree at which the ground heats or cools, and at what magnitude, depends on the timing and duration of the seasonal snow cover (Zhang 2005, Westermann et al. 2011). The insulation effect saturates when the snow thickness reaches 40–50 cm (Zhang 2005; Slater et al. 2017). In addition, snow exerts an important meltwater input (Beniston et al. 2018). With all of the effects taken into consideration, snow cover increases ground temperatures (Smith & Riseborough 1996; Zhang et al. 1997; Osterkamp 2007; Ekici et al. 2015), but the magnitude of the effect on the deeper soil depends on additional factors, such as the ground material, ALT and the presence of moisture (Romanovsky et al. 2010; Throop et al. 2012). In certain regions of discontinuous and sporadic permafrost, the absence of snow cover can be a key factor for permafrost development (Seppälä 1997; Zhang 2005; Smith & Riseborough 2002; Biskaborn et al. 2019). According to Frauenfeld et al. (2004), deep snow during the preceding winter can contribute to the ALT in two ways: causing higher spring and summer moisture and inheriting a shallower freeze depth. Soil temperatures also show pronounced spatial variability related to regional climate-driven snowpack properties, such as density and moisture (Wang et al. 2016).
2.1.2 Climatic sensitivity of permafrost
A consensus has formed that the magnitude of future climate warming will greatly affect the temperature and the extent of permafrost (e.g. AMAP 2017; Biskaborn et al. 2019;
IPCC 2019). Depending on the climate scenario, suitable conditions for permafrost have been predicted to disappear from the present discontinuous zone (assuming medium
8 9 human-induced greenhouse gas emissions) or retreat to encompass mostly high-Arctic
conditions and continental Siberia (high emissions) by 2100 (Slater & Lawrence 2013).
Hence, reducing emissions could help mitigate the impacts of the warming on permafrost.
Chadburn et al. (2017) and Wang et al. (2019) concluded that attaining the targeted 1.5 °C global warming (compared to the pre-industrial period 1850–1900) proposed in the Paris Agreement (UNFCCC 2015) could prevent ~2 × 106 km2 of permafrost from thawing when compared to 2.0 °C warming (cf. Guo & Wang 2017a). Wang et al. (2019) simulated that 1.5 °C warming would be reached as soon as in the 2020s regardless of the used emission trajectory. According to the special report by the Intergovernmental Panel on Climate Change (IPCC 2019), the mean surface air temperature over land areas (excluding oceans) had already reached this limit by 2006–2015. It has already been earlier recognized that the focus should be set for adaptive policies to mitigate the negative outcomes of warming climates (Knight & Harrison 2013; AMAP 2017).
The responsiveness of the ground to climatic conditions depends on its initial thermal state, predominantly the presence or absence of permafrost. In permafrost conditions, the active layer acts as a buffer impeding the translation of atmospheric temperatures to deeper permafrost (Fig. 1, Osterkamp & Romanovsky 1999; Throop et al. 2012; Luo et al. 2016). In non-frozen soils the effect of the climate signal is more direct (Kurylyk et al.
2014; Ekici et al. 2015). Notwithstanding, varying sensitivities exist inside the permafrost domain. Shur and Jorgenson (2007), for example, classified permafrost types based on the interactions of climatic and ecological processes involved in permafrost formation and degradation. According to them, the development of continuous permafrost is mostly driven by climate, while the effects of ecosystem characteristics gain importance towards warmer permafrost regions.
2.1.3 Terrain properties
On local to regional scales, the topography exerts control over the air temperature as a function of elevation, solar radiation and snow transport (Harris et al. 2003). Moreover, the terrain curvature and slope regulate water distribution in the soils (e.g. Etzelmüller et al. 2001), which is inherently dependent on the cohesive and water retaining properties of the soil. The presence of water in the ground either in liquid state or frozen, i.e.
as ground ice, is a key factor affecting the response of permafrost thermal regimes to climatic changes (e.g. Riseborough 1990; Kurylyk et al. 2014). More precisely, permafrost warming is slowed by the higher demand of energy to melt ground ice in the active layer (i.e. latent-heat exchange, Riseborough 1990). The autumnal re-freeze, in turn, is delayed due to the latent heat present in the liquid water in the soil (Romanovsky & Osterkamp 2000; Romanovsky et al. 2010).
The principal heat transfer mode between geothermal and atmospheric heat in permafrost is conduction (Smith & Riseborough 1996). Convective heat transport (by
9
moving water or water vapor) is smaller owing to the limited water movement in frozen soils (e.g. Lachenbruch & Marshall 1986; Harris et al. 2009; Weismüller et al. 2011) yet it is still significant in certain conditions (Kurylyk et al. 2014, Yin et al. 2017; Nicolsky &
Romanovsky 2018). The annual air temperature signal attenuates at shallower depths in soil compared to bedrock due to differences in the thermal conductivity (Smith et al. 2010).
Furthermore, coarse-grained soils have greater conductivity than fine-grained soils (e.g.
Riseborough 1990; Callaghan et al. 2011) and peat is less conductive than mineral soils in all ice/liquid/gas saturation states (Atchley et al. 2016).
Soils with a high organic carbon content have low conductivity in a dry state, whereas when wet they can transfer energy effectively (Kane et al. 2001). Especially frozen, wet organic soils can have a high volumetric ice content and high conductivity, which allow for effective cooling of the ground during cold weather. In the summertime, dry organic soil acts as a weakly conductive insulator that prevents the ground from warming (Kane et al. 2001). Attributed to this asymmetrical heat flow, permafrost may persist in isolated patches with a high organic content outside discontinuous permafrost areas (Shur &
Jorgenson 2007; Harris et al. 2009). The impact of soil organic matter on the thermal diffusivity of soils has been shown to also markedly affect the broad-scale permafrost dynamics (Lawrence et al. 2008; Zhu et al. 2019). Schuur and Mack (2018) suggested that soil organic carbon should be considered as an integrative part of permafrost (along with the temperature and ground ice content) in dictating its response to climate change and further consequences for ecosystems and society.
During the snow-free season, the vegetation layer controls the energy and water exchange between the ground and the atmosphere (Zhang et al. 2003; Ekici et al. 2015).
Vegetation affects the ground thermal regime by intercepting the snowfall and solar radiation (Smith et al. 2010; Woo 2012), reducing the summer heat flux and retaining the insulating snow layer during the winter (Walker et al. 2003; Zhang 2005; Luo et al. 2016), and cooling the ground surface and controlling the water balance through evapotranspiration (Smith & Riseborough 1996, 2002; Jorgenson et al. 2010; Gruber et al. 2017).
In a recent global-scale study De Frenne et al. (2019) demonstrated that differences between the temperatures under forest canopies and adjacent open areas can exceed the magnitude of global warming over the last century. Although the study was conducted in non-permafrost forests (boreal to tropical biomes), the magnitude of the found temperature offsets suggests that future assessments of fine-scale permafrost variability could benefit from considerations of understory microclimates across the vast taigas in Siberia and North America. Local-scale studies have indeed demonstrated that in near- zero MAGT conditions the snow retaining effect (Sladen et al. 2009) or insulating and moisture retaining effects of the vegetation layer per se (Yin et al. 2017) can determine whether permafrost is present or absent. Another local determinant of thermal ground conditions is the presence of open water. Water bodies in permafrost regions induce talik formation (unfrozen layer in permafrost) and can account for ground temperature anomalies of several degrees centigrade in their vicinities (Jorgenson et al. 2010; Woo 2012).
10 11
2.2 Permafrost landforms
Permafrost landforms can be considered to be a subcategory of a wider group of periglacial landforms, which occur also in regions of seasonally frozen ground, and thus do not require permafrost to develop. In this thesis, three iconic landforms encountered in permafrost environments across the Northern Hemisphere were studied. Pingos are intrapermafrost ice-cored mounds that form by two principal processes. Hydrostatic pingos typically form due to the aggradation of ice lenses in low-lying drained lake basins, while more topographical relief is associated with hydraulic pingos that depend on water moving under a hydraulic gradient (Mackay 1973; Jones et al. 2012, Fig. 2a). The majority of pingos occupy circum-Arctic lowlands (Grosse & Jones 2011). Ice-wedge polygons are ubiquitous in unconsolidated soils in the continuous permafrost zone (Bernard- Grand’Maison & Pollard 2018, Fig. 2b). Ice wedges form by cyclical freezing of melt waters in frost cracks that form during winter cold spells (Washburn 1980). Rock glaciers, found in the periglacial belt of all major mountain environments (Barsch 1988), consist of a mixture of ice and poorly sorted debris that flows under gravity due to internal deformation (Barsch 1988; Berthling 2011, Fig. 2c). Multiple rock glacier typologies based on morphology, activity or genesis, for example, have been proposed but often two schools are differentiated based on the origin of a rock glacier. The permafrost creep school (e.g.
Haeberli et al. 2006) assumes that rock glaciers aggregate ice as they form from loose talus or other slope material, whereas the other view is that ice is derived from a glacier that becomes mixed with debris and eventually forms a rock glacier (Whalley & Martin 1992). In this thesis, all types of rock glaciers, apart from debris-covered glaciers (e.g.
Anderson et al. 2018), were considered.
Most studies of permafrost landforms have been conducted on local to regional scales involving aspects such as detailed follow-up studies of single landforms (see Mackay 1972;
Humlum & Christiansen 2008) or descriptive (sometimes statistical) assessments of the observed environments for a set of landforms (see Grosse & Jones 2011; Ran & Liu 2018). Broad-scale efforts are less frequent except Grosse and Jones’s (2011) spatial data analysis of pingos across northern Asia. Recently, regional inventories of rock glaciers, in particular, have become increasingly abundant due to the increased availability of high- resolution remote sensing-based data products and imagery provided by Google Earth, for example (Schmid et al. 2015; Ran & Liu 2018; Du et al. 2019). A comprehensive review of global inventory works concluded that at least 73,000 rock glaciers exist globally (Jones et al. 2018). Grosse and Jones (2011) estimated that all types of pingos considered there are at least 11,000 pingos on Earth. Broad-scale modelling assessments, however, have been hindered by the lack of homogenized datasets of landform occurrences. Moreover, (predictive) distribution modelling studies of pingos, ice-wedge polygons and rock glaciers, are lacking apart from a few regional-scale studies (e.g. Brenning et al. 2007; Marcer et al.
2017; O’Neill et al. 2019).
11
Figure 2. Examples of the studied landforms in satellite imagery. Two large pingos (a) on the Arctic Ocean coast near Tuktoyaktuk, Canada (69.399 °N, 133.079 °W, Image © Google Earth, CNES/Airbus), ice-wedge polygons (b) at various stages of degradation on the Yamal Peninsula (72.354 °N 72.555 °E, Image © Google Earth, Maxar Technologies), and rock glaciers (c) on Disko Island, Greenland (69.417 °N, 53.926 °W, Image © Google Earth, Maxar Technologies).
12 12
2.3 Human activity and permafrost
The adverse effects of permafrost aggradation and degradation on buildings or transportation infrastructure are by no means a new phenomenon. Human settlements and the utilization of natural resources in the Arctic have long been affected by ground ice dynamics, and adaptation measures been developed historically (Mackay 1972; Péwé 1979;
Koutaniemi 1985; Shiklomanov 2005). Recently, the intensifying socio-economic activity in the Arctic has evolved into a growing concern for the local and global consequences that rapidly warming permafrost may have in the near future (Callaghan et al. 2011; Vincent et al. 2017; IPCC 2019). This amusingly illustrates the development of cryospheric science during the past decades, given that in the 1960s research of underground ice according to Mackay (1972: 12) was denoted by many as a “scientific luxury, interesting, but of little concern to the affairs of modern man”. This view, according to Mackay (1972), rapidly changed after the early oil extraction industry was faced with the problems associated with melting ground ice. The vulnerability of the infrastructure to thawing permafrost, however, had been recognized in early literature in Russian (Sumgin 1927) and Alaskan contexts (Muller 1947).
Notwithstanding the remoteness and harsh climate, the permafrost regions in the Northern Hemisphere are home to millions of people and possess vast amounts of natural resources, such as hydrocarbon deposits (Streletskiy & Shiklomanov 2016; Badina 2017;
Streletskiy et al. 2019). Assessing the wide-ranging implications of change in the Arctic necessitates an integrated approach that considers the consequences and interrelationships between multiple natural and societal aspects (Vincent et al. 2017; Schuur & Mack 2018). In this thesis, I adopt a systematic approach, in which permafrost is considered a regulator of the impacts of climate change to infrastructure. The system, however, has other multidirectional relationships, e.g., the direct effects of the infrastructure on local permafrost degradation or those of biogeochemical feedback mechanisms (e.g. greenhouse gas emissions from thawing permafrost) on the global climate, which are beyond the analytical scope of the thesis.
Despite the growing global relevance of the Arctic and abundant Arctic-wide, local evidence of permafrost degradation-related damage to infrastructure (e.g. Grebenets et al. 2012; Doré et al. 2016; Shiklomanov et al. 2017), the spatial distribution of hazards on a circumpolar scale is insufficiently known. Among the first mapping efforts, Nelson et al. (2001, 2002) formulated a geohazard index which is applicable to assessing ground settlement due to an increasing ALT across the Northern Hemisphere. The settlement index and its remakes (Anisimov & Reneva 2006; Guo & Sun 2015; Guo & Wang 2017a), however, only considered one type of geohazard and were highly generalized and coarse in their spatial resolution. Thus, they were of limited value to address within-region variability concerning the hazard potential. In addition, permafrost degradation-related risks to specific infrastructure elements or population centers have not been quantified on a circumpolar scale.
13
This thesis focuses on the Northern Hemisphere land areas north of the 30th latitude (Fig. 3). The main focus is on the permafrost domain but non-permafrost regions, parts of them characterized by seasonal freezing (Zhang et al. 2003), are also addressed for methodological (see Papers III-IV) and comparative (Paper I) reasons. The extensive areas of high-altitude permafrost in the Tibetan Plateau and in other mid-latitude mountain areas are here considered as parts of the circumpolar permafrost domain. Permafrost is typically classified based on its lateral extent. The zone of continuous permafrost, by definition affecting more than 90% of an area (Zhang et al. 1999), characterizes the coldest areas most extensively found in Russia, Canada and Alaska (Fig. 3). In northern Alaska, the permafrost thickness can reach depths of ~500 and up to 1,500 m in northern Sibe- ria (Washburn 1980; Bodri & Cermak 2007). As environmental conditions become less suitable for permafrost, its extent becomes discontinuous (50–90% cover) and sporadic (10–50%). Outside these zones isolated patches of permafrost (< 10%) occur in locations with suitable microclimates (e.g. high elevations, ice caves) or soil/vegetation conditions that allow for permafrost to persist under warmer conditions (e.g. palsa mires, Seppälä 1997; Luoto & Seppälä 2002; Shur & Jorgenson 2007). All zones considered; permafrost covers ~23 x 106 km2 (24%) of the exposed land areas in the Northern Hemisphere (Zhang et al. 1999).
3 Study area
14 14
Figure 3. The Northern Hemisphere study area north of 30th latitude and the used observational datasets. Locations for the compiled mean annual ground temperature (MAGT, n = 797) and active-layer thickness (ALT, n = 303) observations are displayed in panel a. The permafrost landform occurrences included 9,709 pingos, 861 ice-wedge polygons and 4,035 rock glaciers (b). The permafrost domain is shaded by its extent after Brown et al. (2002).
15
4.1 Response data
4.1.1 Observations of mean annual ground temperature and active-layer thickness
The MAGT and ALT observations (Fig. 3a) used in Papers I, III and IV were compiled from previous databases, research articles and maps (Supplementary Tables 1–2 in Paper III). In Paper I, the MAGT at or below 0 °C (representative of permafrost) and MAGT above 0 °C (non-permafrost) were analysed separately. The preconditions for the included MAGT and ALT observations implied that they were 1) not disturbed by geothermal or anthropogenic heat sources, recent fires or large water bodies in the immediate vicinities, 2) recorded during the study period (2000–2014), and 3) had an adequate locational accuracy in order to comply with the used geospatial data at a 30 arc-second resolution. If more than one observation occupied the same grid cell, a median value was used. Additional dataset-wise steps were taken to ensure that the field observations were comparable across the study area and are described below.
The MAGT can be determined for any observed depth using temperature measurements from boreholes, for example. The temporal resolution of the used ground temperature data varied between year-round hourly records to single once-in-a-year observations.
Continuous records were averaged over a year, and when single observations were used it was ensured that the given depth was not affected by intra-annual temperature fluctuations.
To address spurious measurements, ground temperatures at or near the DZAA (Section 2.1) in near-surface permafrost (on average 12.5 meters below the ground surface) were utilized. This delineation was necessary to balance between 1) filtering out inter-annual fluctuations occurring in the top soil layers (Romanovsky et al. 2002), 2) ensuring that the ground thermal regime is in balance with current climatic conditions on a decadal time scale (Romanovsky et al. 2007), and 3) thus being responsive to projected changes in the climate in the studied future periods (2041–2060 and 2061–2080). It is important to note that even though deep permafrost occurring at depths of several tens to hundreds of meters can persist in a changing climate (Lachenbruch & Marshall 1986; Huang et al. 2000), near-surface layers are anticipated to show quick and significant responses to even temporary decadal-scale air temperature increases (Guo & Wang 2017b; Zhang et al. 2018a).
Field measurements of active-layer thickness are done either by physically probing the ground at the time of maximum thaw or reading the exact depth of the thaw from year-round installed thaw tubes (Brown et al. 2000). In addition, the thaw depth can be inferred from soil temperature profiles; the ALT value is a product of an interpolation between the nearest two measurements above and below the depth of the maximum thaw
4 Materials and methods
16 17 (Nelson & Hinkel 2004). Comparability between the ALT measurements derived from
different methods and at different locations was achieved by using values that represent the maximal annual thaw recorded during late summer to autumn (Hinkel & Nelson 2003;
Bonnaventure & Lamoureux 2013). Figure 4 presents an overview of the performed research from data preparation to statistical analyses, geohazard index formulation and quantifications of the infrastructure at risk.
Figure 4. The study design scheme presenting the used data, methods (light grey boxes) and outcomes (dark grey ovals) of the performed analyses. To ensure globally comparable observational data on the mean annual ground temperature (MAGT), active-layer thickness (ALT), permafrost landforms and infrastructure and population data, criterion frameworks were developed and applied (see text). Explorative analyses included assessments of sample statistics and bivariate correlations in the modelling data. Statistical modelling was performed applying ensembles of four modelling techniques. The produced predictions of the MAGT and ALT, together with additional geospatial data on geohazard-affecting factors, were applied in geohazard index formulation. Finally, the geohazard indices were employed in overlay analyses, in which the susceptibility of the infrastructure and population to near-surface permafrost degradation-related geohazards was quantified.
4.1.2 Permafrost landform observations
In order to perform a circumpolar modelling of permafrost landform distributions, datasets of pingo, ice-wedge polygon and rock glacier occurrences (Fig. 3b) were compiled from available sources (Supplementary Table 3 in Paper II). In the compilation, the primary types of landforms were grouped together, although in some cases, the geomorphic processes behind their formation may differ (e.g. Harris 1981; Knight et al. 2019). Similarly to MAGT and ALT data, only one observation was assigned for each grid cell to avoid pseudoreplication (Hurlbert 1984). By aiming to include all the documented occurrence areas I sought to minimize sampling biases, e.g., observations clustered only in well-studied and accessible regions. However, extensive under-sampled regions remained. Overall, these data are suggested to be the most geographically comprehensive spatial data compilations of pingos, ice-wedge polygons and rock glaciers covering the Northern Hemisphere.
Method-wise, modelling geomorphic landforms (presence-absence) and MAGT and ALT (continuous response) is similar, but some geomorphological discourses can be
17
attached to the study of landforms. First, landforms are simplifications of their natural environment where they are highly variable in form and occur in continuums of landscape elements. Thus, they are a product of classification. For example, among the contested discourse about rock glacier typologies (e.g. Barsch 1988; Whalley & Martin 1992) there is still some ongoing discussion on the continuum of glacier–debris-covered glacier–rock glacier development (e.g. Anderson et al. 2018; Jones et al. 2019a; Knight et al. 2019). So far, no canonical thresholds to delineate rock glacier types exist, although new classification schemes based on the surface movement dynamics, for example, have been proposed (Knight 2019). Similar discussions have lingered around pingos, which are usually thought to form because of two different groundwater pressure mechanisms (Mackay 1973, 1998). A process-based classification in practice, however, is often not feasible as both mechanisms can occur simultaneously (Worsley & Gurney 1996; Gurney 1998).
To confront such hindrances, Paper II assumed a concept of equifinality whereby similar landforms are thought to result from a varying set of processes and initial conditions (Slaymaker 2004). Such generalizations were needed in order to achieve comprehensive modelling datasets on a circumpolar scale, i.e., to cover the entire gradient of environmental conditions occupied by known landform occurrences. Another central concept to the study of future distributions is that of uniformitarianism. James Hutton proposed that the Earth’s geological history can be explained in terms of natural forces acting today, and further, that the same applies for the future (Hutton 1788). In this thesis, the view was reflected in an assumption that past climate-induced shifts in landform distributions, as demonstrated by the remnants of pingos and ice-wedge patterns across currently temperate climate regions (Vanderberghe & Pissart 1993; Mackay 1998;
Vandenberghe et al. 2014), would be anticipated in future changing climates.
4.2 Geospatial data
Numerous web-based data depositories and data infrastructures have made geospatial data increasingly available (Dowman & Reuter 2017). In this thesis, geospatial data constituted digital data layers (raster and vector) that could be processed with geographical information system (GIS) software; herein primarily ArcGIS (ESRI 2015), R (R Core Team 2015) and the System for Automated Geoscientific Analyses (SAGA GIS, Conrad et al. 2015). Given the marked reliance of the statistical analyses on correlations in the data (Marmion et al.
2009), care had to be taken to choose the most appropriate predictors for each response (Austin et al. 2006; Hjort & Luoto 2013). Based on the literature, the aim was to involve all physically relevant predictors of sufficient spatial resolution and coverage (Table 1). Apart from ground ice content data (Brown et al. 2002), originally at 12.5 km spatial resolution, all predictors had a native resolution of 30 arc seconds (< 1 km2 grid cell size) or finer, and were resampled to 30 arc second resolution prior analyses. The geographical extent of all the predictors was limited at the 30th latitude.
18 19
Table 1. The geospatial and infrastructure data used in individual papers. The abbreviated dataset acronyms are: Global Meteorological Forcing Dataset (GMFD), United States Geological Survey (USGS) Shuttle Radar Topography Mission (STRM) Digital Elevation Model (DEM), Global Multi-resolution Terrain Elevation Data (GMTED), International Permafrost Association (IPA), European Space Agency Climate Change Initiative (ESA CCI) and Moderate Resolution Imaging Spectroradiometer (MODIS). Original datasetDerived parametersUsed in PapersOriginal data citation WorldClim v1.4
Freezing degree-days I–IVHijmans et al. 2005Thawing degree-days Snowfall Rainfall GMFDClimate reanalysis adjustment parameters
I, III– IV
Sheffield et al. 2006 USGS SRTM DEM
Potential incident solar radiation (after McCune & Keon 2002)
I, III– IV
United States Geological Survey 2004 Slope gradientIII–IV GMTED
Potential incident solar radiation (after McCune & Keon 2002) IIDanielson & Gesch 2011Slope gradient Topographical Wetness Index (after Böhner & Selige 2006) SoilGrids1kmSoil organic carbon content III–IVHengl et al. 2014Coarse sediment content Fine sediments content
19
SoilGrids250mSoil organic carbon content I–IIHengl et al. 2017Coarse sediment content Fine sediments content Soil and sediment deposit thicknessIII–IVPelletier et al. 2016 IPA Volumetric ground ice contentIII–IVBrown et al. 2002 ESA CCI Water Bodies v4Coverage of water bodiesII–IVDefourny 2016 MODIS TerraNormalized Difference Vegetation IndexIDidan 2015 OpenStreetMap
Roads IVOpenStreetMap Contributors 2016, www.openstreemap.org
Railways Oil and gas pipelines Buildings Industrial areas Populated settlements OurAirportsAirports and airfieldsIVOurAirports.com RosnedraHydrocarbon extraction areas in RussiaIVgis.sobr.geosys.ru Gridded
Center for International Earth Science Population of Human populationIVInformation Network (CIESIN) the World (2016) (GPW v4)
20 21 4.2.1 Current and future climates
Attributed to the seasonal asymmetries in the response between atmospheric and ground thermal regimes (see Section 2.1.1), winter and summer air temperatures and precipitation were considered separately. Several previous studies have demonstrated that indices representing the length or magnitude of the thawing and freezing seasons are often more suitable for permafrost modelling than mean annual air temperature (e.g. Zhang et al. 1997; Harris et al. 2009; Smith et al. 2009). Four climatic parameters;
freezing and thawing degree days (FDD and TDD, °C-days), and the snow- and rainfall estimated from monthly air temperature averages were computed from gridded data on interpolated monthly climate surfaces in the WorldClim database (Hijmans et al. 2005) for baseline periods of 2000–2014 (Papers I, III–IV) and 1950–2000 (Paper II). For the former case, the native period (1950–2000) of the WorldClim data had to be adjusted (see Aalto et al. 2018a) using the Global Meteorological Forcing Dataset for land surface modelling (Sheffield et al. 2006) to match the period (2000–2014) that MAGT and ALT observations were representative of.
The climatic sensitivity of permafrost was assessed by estimating the influence of changing climatic parameters on the model outputs (i.e. the predicted MAGT and ALT, permafrost landform distributions, and spatio-temporal patterns in projected geohazards) (Fig. 4). Future climates were based on climate and Earth system models from the Coupled Model Intercomparison Project (CMIP5, Taylor et al. 2012). Different trajectories of human-induced climate change were taken into account using the representative concentration pathways (RCPs, van Vuuren et al. 2011). RCPs represent the estimated radiative forcing values by 2100 based on human-induced greenhouse gas emissions; according to the most optimistic (RCP2.6) pathway, emissions peak in the 2020s while the ‘business-as-usual’
pathway (RCP8.5) assumes a constant increase (van Vuuren et al. 2011). In this thesis, two future periods were considered; mid-century (2041–2060) and late-century (2061–2080).
To address a broad spectrum of model responses and associated uncertainty (Thuiller et al.
2019), multiple emission trajectories, namely RCP2.6, RCP4.5 and RCP8.5, were included in the assessment of permafrost degradation-related hazards to the infrastructure (Papers III–IV). In the exploration of the climate change effects on permafrost landforms (Paper II), RCP4.5 and RCP8.5 were considered.
4.2.2 Terrain properties
Digital elevation models (DEMs) were the first-order source for topographical predictors.
As discussed in section 2.1.3, the topography regulates the local air temperature and soil moisture conditions, for example (Etzelmüller et al. 2001). It is important to note, that the model fit and predictive performance are influenced by the resolution of the geospatial data layers (Yates et al. 2018). DEM-derived predictors at a 30 arc-second resolution (~1
21
km2) were here assumed to represent terrain properties on scales which are relevant to the local variability of the studied responses on a circumpolar scale. Notwithstanding, finer-scale variations in micro-climatic, soil and hydrological conditions especially in heterogeneous topographies undoubtedly exist (e.g. Hoelzle et al. 2001; Etzelmüller 2013;
Fiddes et al. 2015; Aalto et al. 2018b).
Solar radiation input was computed using the parameterization by McCune & Keon (2002). Based on a DEM-derived slope, latitude and aspect, the method yielded an annual estimate of the potential incident solar radiation (PISR, MJ cm-2 y-1). The slope gradient, computed using the ArcGIS Spatial Analyst extension, was used as an independent factor in geohazard formulation and in Paper II. The topographical wetness index (TWI), used in Paper II, was computed in SAGA GIS with the SAGA Wetness Index tool (Böhner et al. 2002). In addition to the slope, it involved a computation of the specific catchment area (Böhner & Selige 2006). The index represents the accumulation potential of water in a grid cell based on its position in the catchment area.
Soil properties were derived from data layers in the SoilGrids database (Hengl et al.
2014, 2017). The contents of soil organic carbon (SOC, g kg-1), coarse sediments (coarse fragments > 2 mm, %) and fine sediments (sum of clay and silt, ≤ 50 µm, %) were averaged over seven depth intervals from the ground surface to a depth of 200 cm. In Papers III and IV, SOC data provided for the depth of 60–100 cm was used. The geohazard index parameters in Paper III involved an estimation of soil and sediment thickness, for which the gridded data by Pelletier et al. (2016) was used. The classic “Circum-Arctic Map of Permafrost and Ground Ice Conditions” by Brown et al. (2002) provides the only currently available circumpolar spatial data on the ground ice content. The classified volumetric ground ice content zonation in this data was used in geohazard formulation. The potential contribution of water bodies to ground thermal regimes was taken into account in Papers II–IV using remote sensing data (Defourny 2016). Finally, the effects of vegetation cover on the MAGT and ALT were assessed by computing a normalized difference vegetation index (NDVI, Didan 2015) averaged over summer months (June to August) for the 2000–
2014 period using the Moderate Resolution Imaging Spectroradiometer (MODIS) data.
4.2.3 Infrastructure data
Prior circumpolar assessments of permafrost degradation-related geohazards have not explicitly determined the amount of infrastructure at risk. This is due partly to their coarse spatial resolution but also to the lack of available globally coherent data on infrastructure elements. In Paper IV, sub-square-kilometre spatial resolution of mapped geohazards required spatially accurate data on the studied infrastructure, which was compiled from available databases (Table 1). The included infrastructure elements were chosen based on their relevancy to the human and industrial utilization of the permafrost regions.
Linear features consisted of transportation infrastructure; roads, railways and pipelines,
22 23 whereas airports and populated settlements were included as point locations, and buildings,
industrial areas and hydrocarbon extraction areas as polygon footprints. Most of the data were derived from the national excerpts of OpenStreetMap database (OpenStreetMap contributors 2016) acquired from geofabrik.de. Some of the included features were reclassified in order to reduce the risk of data quality discrepancies across the study area. For example, only the five most important types of roads were included in order to alleviate the spatially imbalanced data completeness, i.e. developed countries and urban areas have a higher mapping density than less developed and rural areas (Barrington-Leigh
& Millard-Ball 2017). In addition, raster data on census-based human population for the year 2015 (Center for International Earth Science Information Network 2016) was used to characterize the human distribution across the permafrost domain.
4.3 Statistical modelling
In statistical modelling, responses (here MAGT, ALT and permafrost landform presence/
absence) are related to predictors (environmental conditions represented by geospatial data) through correlative relationships (e.g. Guisan & Zimmermann 2000; Hoelzle et al.
2001; Yates et al. 2018). More precisely, the models are calibrated by fitting a function (here using regression or machine-learning algorithms) between field observation and the values for multiple predictors at the corresponding location. In this thesis, multivariate statistical models were implemented both to explain complex process-environment relationships and thereby provide insights into the functioning of Earth surface systems on a circumpolar scale, and to predict permafrost parameters in space and time.
Based on the explorative analyses, moderately strong (> r |0.7|) bivariate correlations potentially introducing a multicollinearity effect (Dormann et al. 2013) occurred between certain predictors in a few modelling datasets, e.g. between air temperature and solar radiation (Fig. 2 in Paper I), rainfall and TDD, and TWI and slope (Supplementary Figure 3 in Paper II). Omitting physically relevant predictors with distinct hypothesised effects on the responses, however, was not desirable. Notwithstanding, the possible effects of multicollinearity had to be carefully considered in interpretations of the results (Dormann et al. 2013).
Earth surface systems and geomorphic processes often have a nonlinear nature, i.e. the outputs of a system are not proportional to the inputs across their entire range (Phillips 2006). Nonlinearities can be examined with response curves, which can yield insights into the often-complex process-environment relationships and facilitate theoretical discussions (Hjort & Luoto 2011). Visualizing the responses also brings transparency to the model evaluation and interpretation (Elith et al. 2005) by allowing assessments of realism between the found correlative relationships (Austin et al. 2006). Here, response curves were used to assess the shape and direction of the responses of MAGT and ALT (Paper I) and permafrost landforms (Paper II) to environmental predictors. In addition,
23
Paper I included an analysis of the effect size that could be used to assess the magnitude of each predictor in the units of the response (°C for MAGT, cm for ALT). To allow further insights, Papers I and II included analyses of variable importance (Breiman 2001).
In this procedure, the correlation (Pearson’s r) is computed between a model fitted with all the environmental predictors and another model where one predictor is randomized.
The procedure is repeated to randomize each predictor in turn. The higher the variable importance score, the higher individual contribution a predictor has to a response (Thuiller et al. 2009). Predictive distribution modelling was a central part of Papers II–IV. Statistical ensemble forecasting of MAGT and ALT (Papers III–IV) was applied to derive the current permafrost extent (2000–2014, modelled MAGT ≤ 0 °C) and ALT distribution, and to assess corresponding changes in future conditions (2041–2060 and 2061–2080) under climate-change scenarios (Section 4.2.1). In paper II, the current and future distributions of the studied permafrost landforms were predicted using a similar approach.
All the analyses in this thesis were performed using four statistical modelling techniques (see Papers I–IV and Aalto et al. 2017, 2018a for details). Generalized linear models (GLMs, Nelder & Wedderburn 1972) and generalized additive models (GAMs, Hastie
& Tibshirani 1986) have been frequently applied in predictive spatial modelling contexts (statistical distribution modelling) (Guisan et al. 2002). Multivariate statistical modelling in geomorphological research began to emerge in the late 20th century (see Luoto & Hjort 2005). More recently, machine learning-based methods have gained ground in the fields of distribution modelling (Luoto & Hjort 2005; Hao et al. 2019). Here a generalized boosting method (GBM, Elith et al. 2008) and random forest (RF, Breiman 2001) algorithms employing machine-learning were used. The final ALT modelling in Papers III–IV was based only on GLM.
The multi-model approach comes with both advantages and challenges. The inclusion of machine learning to accompany conventional statistical methods brings efficient alternatives for model selection, tuning and evaluation (Elith et al. 2008; Marmion et al.
2008). For example, GBM allows for building models that iteratively concentrate on improving the predictions of initially weakly explained cases across the dataspace, thereby reducing bias and variance (Friedman et al. 2000). Additionally, in contrast to GLM and GAM, machine learning methods can automatically select relevant predictors and identify and model interactions among them (Friedman et al. 2000; Elith et al. 2005, 2008; Thuiller et al. 2009). GLM and GAM are additive methods, and here they were applied by simply summing up the contributions of the model terms (fitted predictors), without taking into account the interactions among them (Elith et al. 2005). Hence, owing to the different abilities of the models to handle aspects including collinearity, spatial autocorrelation or nonlinearity, the models may perform differently with the used environmental data (Marmion et al. 2009; Zhu & Peterson 2017).
The main reasons for employing ensemble and model-averaging techniques were to account for inter-model variability and reduce the uncertainty involved in choosing a single modelling method (Araújo & New 2007, Marmion et al. 2009; Thuiller et al. 2009). These
